networks.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import init
from torchvision import models
import os
import numpy as np

class Options:
    def __init__(self):
        # Default values
        self.fine_height = 256
        self.fine_width = 192
        self.grid_size = 5
        self.use_dropout = False
        self.input_nc = 22
        self.input_nc_B = 1  
        self.tom_input_nc = 26  
        self.tom_output_nc = 4  

def weights_init_normal(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        init.normal_(m.weight.data, 0.0, 0.02)
    elif classname.find('Linear') != -1:
        init.normal(m.weight.data, 0.0, 0.02)
    elif classname.find('BatchNorm2d') != -1:
        init.normal_(m.weight.data, 1.0, 0.02)
        init.constant_(m.bias.data, 0.0)

def init_weights(net, init_type='normal'):
    print('initialization method [%s]' % init_type)
    net.apply(weights_init_normal)

class FeatureExtraction(nn.Module):
    def __init__(self, input_nc, ngf=64, n_layers=3, norm_layer=nn.BatchNorm2d, use_dropout=False):
        super(FeatureExtraction, self).__init__()
        downconv = nn.Conv2d(input_nc, ngf, kernel_size=4, stride=2, padding=1)
        model = [downconv, nn.ReLU(True), norm_layer(ngf)]
        for i in range(n_layers):
            in_ngf = 2**i * ngf if 2**i * ngf < 512 else 512
            out_ngf = 2**(i+1) * ngf if 2**i * ngf < 512 else 512
            downconv = nn.Conv2d(in_ngf, out_ngf, kernel_size=4, stride=2, padding=1)
            model += [downconv, nn.ReLU(True), norm_layer(out_ngf)]
        model += [nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1), nn.ReLU(True)]
        model += [norm_layer(512)]
        model += [nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1), nn.ReLU(True)]
        self.model = nn.Sequential(*model)
        init_weights(self.model)

class FeatureL2Norm(nn.Module):
    def __init__(self):
        super(FeatureL2Norm, self).__init__()

    def forward(self, feature):
        epsilon = 1e-6
        norm = torch.pow(torch.sum(torch.pow(feature, 2), 1) + epsilon, 0.5).unsqueeze(1).expand_as(feature)
        return torch.div(feature, norm)

class FeatureCorrelation(nn.Module):
    def __init__(self):
        super(FeatureCorrelation, self).__init__()

    def forward(self, feature_A, feature_B):
        b, c, h, w = feature_A.size()
        feature_A = feature_A.transpose(2, 3).contiguous().view(b, c, h*w)
        feature_B = feature_B.view(b, c, h*w).transpose(1, 2)
        feature_mul = torch.bmm(feature_B, feature_A)
        return feature_mul.view(b, h, w, h*w).transpose(2, 3).transpose(1, 2)

class FeatureRegression(nn.Module):
    def __init__(self, input_nc=512, output_dim=6):
        super(FeatureRegression, self).__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(input_nc, 512, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(512),
            nn.ReLU(inplace=True),
            nn.Conv2d(512, 256, kernel_size=4, stride=2, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU(inplace=True),
            nn.Conv2d(256, 128, kernel_size=3, padding=1),
            nn.BatchNorm2d(128),
            nn.ReLU(inplace=True),
            nn.Conv2d(128, 64, kernel_size=3, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
        )
        self.linear = nn.Linear(64 * 4 * 3, output_dim)
        self.tanh = nn.Tanh()

    def forward(self, x):
        x = self.conv(x)
        x = x.contiguous().view(x.size(0), -1)
        x = self.linear(x)
        return self.tanh(x)

class TpsGridGen(nn.Module):
    def __init__(self, out_h=256, out_w=192, grid_size=5):
        super(TpsGridGen, self).__init__()
        self.out_h, self.out_w = out_h, out_w
        self.grid_size = grid_size
        
        # Create grid
        axis_coords = np.linspace(-1, 1, grid_size)
        self.N = grid_size * grid_size
        P_Y, P_X = np.meshgrid(axis_coords, axis_coords)
        P_X = torch.FloatTensor(P_X.reshape(-1, 1))
        P_Y = torch.FloatTensor(P_Y.reshape(-1, 1))
        self.P_X_base = P_X.clone()
        self.P_Y_base = P_Y.clone()
        self.Li = self.compute_L_inverse(P_X, P_Y).unsqueeze(0)
        
        # Grid for interpolation
        grid_X, grid_Y = np.meshgrid(np.linspace(-1, 1, out_w), np.linspace(-1, 1, out_h))
        self.grid_X = torch.FloatTensor(grid_X).unsqueeze(0).unsqueeze(3)
        self.grid_Y = torch.FloatTensor(grid_Y).unsqueeze(0).unsqueeze(3)

    def compute_L_inverse(self, X, Y):
        N = X.size()[0]
        Xmat, Ymat = X.expand(N, N), Y.expand(N, N)
        P_dist_squared = torch.pow(Xmat-Xmat.transpose(0, 1), 2) + torch.pow(Ymat-Ymat.transpose(0, 1), 2)
        P_dist_squared[P_dist_squared == 0] = 1
        K = torch.mul(P_dist_squared, torch.log(P_dist_squared))
        O = torch.FloatTensor(N, 1).fill_(1)
        Z = torch.FloatTensor(3, 3).fill_(0)
        P = torch.cat((O, X, Y), 1)
        L = torch.cat((torch.cat((K, P), 1), torch.cat((P.transpose(0, 1), Z), 1)), 0)
        return torch.inverse(L)

    def forward(self, theta):
        theta = theta.contiguous()
        batch_size = theta.size()[0]
        
        # Split theta into point coordinates
        Q_X = theta[:, :self.N].contiguous().view(batch_size, self.N, 1)
        Q_Y = theta[:, self.N:].contiguous().view(batch_size, self.N, 1)
        Q_X = Q_X + self.P_X_base.expand_as(Q_X)
        Q_Y = Q_Y + self.P_Y_base.expand_as(Q_Y)
        
        # Compute weights
        W_X, W_Y = self.apply_theta(Q_X, Q_Y)
        
        # Calculate transformed grid
        points_X, points_Y = self.transform_points(W_X, W_Y)
        return torch.cat((points_X, points_Y), 3)

class GMM(nn.Module):
    def __init__(self, opt=None):
        super(GMM, self).__init__()
        if opt is None:
            opt = Options()
            
        self.extractionA = FeatureExtraction(opt.input_nc)
        self.extractionB = FeatureExtraction(opt.input_nc_B)
        self.l2norm = FeatureL2Norm()
        self.correlation = FeatureCorrelation()
        self.regression = FeatureRegression(input_nc=192, output_dim=2*opt.grid_size**2)
        self.gridGen = TpsGridGen(opt.fine_height, opt.fine_width, opt.grid_size)

    def forward(self, inputA, inputB):
        featureA = self.extractionA(inputA)
        featureB = self.extractionB(inputB)
        featureA = self.l2norm(featureA)
        featureB = self.l2norm(featureB)
        correlation = self.correlation(featureA, featureB)
        theta = self.regression(correlation)
        grid = self.gridGen(theta)
        return grid, theta

class UnetGenerator(nn.Module):
    def __init__(self, input_nc, output_nc, num_downs, ngf=64, norm_layer=nn.InstanceNorm2d):
        super(UnetGenerator, self).__init__()
        unet_block = UnetSkipConnectionBlock(
            ngf * 8, ngf * 8, input_nc=None, submodule=None, norm_layer=norm_layer, innermost=True)
        
        for _ in range(num_downs - 5):
            unet_block = UnetSkipConnectionBlock(
                ngf * 8, ngf * 8, input_nc=None, submodule=unet_block, norm_layer=norm_layer)
                
        unet_block = UnetSkipConnectionBlock(ngf * 4, ngf * 8, input_nc=None, submodule=unet_block, norm_layer=norm_layer)
        unet_block = UnetSkipConnectionBlock(ngf * 2, ngf * 4, input_nc=None, submodule=unet_block, norm_layer=norm_layer)
        unet_block = UnetSkipConnectionBlock(ngf, ngf * 2, input_nc=None, submodule=unet_block, norm_layer=norm_layer)
        
        self.model = UnetSkipConnectionBlock(
            output_nc, ngf, input_nc=input_nc, submodule=unet_block, outermost=True, norm_layer=norm_layer)

    def forward(self, input):
        return self.model(input)

class UnetSkipConnectionBlock(nn.Module):
    def __init__(self, outer_nc, inner_nc, input_nc=None, submodule=None, 
                 outermost=False, innermost=False, norm_layer=nn.InstanceNorm2d):
        super(UnetSkipConnectionBlock, self).__init__()
        self.outermost = outermost
        use_bias = norm_layer == nn.InstanceNorm2d
        
        if input_nc is None:
            input_nc = outer_nc
            
        downconv = nn.Conv2d(input_nc, inner_nc, kernel_size=4, stride=2, padding=1, bias=use_bias)
        downrelu = nn.LeakyReLU(0.2, True)
        downnorm = norm_layer(inner_nc)
        uprelu = nn.ReLU(True)
        upnorm = norm_layer(outer_nc)

        if outermost:
            upconv = nn.ConvTranspose2d(inner_nc * 2, outer_nc, kernel_size=4, stride=2, padding=1)
            down = [downconv]
            up = [uprelu, upconv, nn.Tanh()]
            model = down + [submodule] + up
        elif innermost:
            upconv = nn.ConvTranspose2d(inner_nc, outer_nc, kernel_size=4, stride=2, padding=1, bias=use_bias)
            down = [downrelu, downconv]
            up = [uprelu, upconv, upnorm]
            model = down + up
        else:
            upconv = nn.ConvTranspose2d(inner_nc * 2, outer_nc, kernel_size=4, stride=2, padding=1, bias=use_bias)
            down = [downrelu, downconv, downnorm]
            up = [uprelu, upconv, upnorm]
            model = down + [submodule] + up

        self.model = nn.Sequential(*model)

    def forward(self, x):
        if self.outermost:
            return self.model(x)
        else:
            return torch.cat([x, self.model(x)], 1)

class TOM(nn.Module):
    """ Try-On Module """
    def __init__(self, opt=None):
        super(TOM, self).__init__()
        if opt is None:
            opt = Options()
        
        # Input: [agnostic(3) + warped_design(3) + warped_mask(1) + features(19)] = 26 channels
        self.unet = UnetGenerator(
            input_nc=opt.tom_input_nc,
            output_nc=opt.tom_output_nc,  # [rendered(3) + mask(1)]
            num_downs=6,
            norm_layer=nn.InstanceNorm2d
        )

    def forward(self, x):
        output = self.unet(x)
        p_rendered, m_composite = torch.split(output, [3, 1], dim=1)
        p_rendered = torch.tanh(p_rendered)
        m_composite = torch.sigmoid(m_composite)
        return p_rendered, m_composite

def save_checkpoint(model, save_path):
    if not os.path.exists(os.path.dirname(save_path)):
        os.makedirs(os.path.dirname(save_path))
    torch.save(model.state_dict(), save_path)

def load_checkpoint(model, checkpoint_path, strict=True):
    if not os.path.exists(checkpoint_path):
        raise FileNotFoundError(f"Checkpoint file not found: {checkpoint_path}")
    
    state_dict = torch.load(checkpoint_path, map_location=torch.device('cpu'))
    model.load_state_dict(state_dict, strict=strict)